Graphical Models. Andrea Passerini Statistical relational learning. Graphical Models

Transcription

1 Andrea Passerini Statistical relational learning

2 Probability distributions Bernoulli distribution Two possible values (outcomes): 1 (success), 0 (failure). Parameters: p probability of success. Probability mass function: { p if x = 1 P(x; p) = 1 p if x = 0 Example: tossing a coin Head (success) and tail (failure) possible outcomes p is probability of head

3 Probability distributions Multinomial distribution (one sample) Models the probability of a certain outcome for an event with m possible outcomes {v 1,..., v m } Parameters: p 1,..., p m probability of each outcome Probability mass function: P(v i ; p 1,..., p m ) = p i Tossing a dice m is the number of faces p i is probability of obtaining face i

4 Probability distributions Gaussian (or normal) distribution Bell-shaped curve. Parameters: µ mean, σ 2 variance. Probability density function: p(x; µ, σ) = 1 (x µ)2 exp 2πσ 2σ 2 Standard normal distribution: N(0, 1) Standardization of a normal distribution N(µ, σ 2 ) z = x µ σ

5 Conditional probabilities conditional probability probability of x once y is observed P(x y) = P(x, y) P(y) statistical independence variables X and Y are statistical independent iff implying: P(x, y) = P(x)P(y) P(x y) = P(x) P(y x) = P(y)

6 Basic rules law of total probability The marginal distribution of a variable is obtained from a joint distribution summing over all possible values of the other variable (sum rule) P(x) = y Y P(x, y) P(y) = x X P(x, y) product rule conditional probability definition implies that P(x, y) = P(x y)p(y) = P(y x)p(x) Bayes rule P(y x) = P(x y)p(y) P(x)

7 Playing with probabilities Use rules! Basic rules allow to model a certain probability given knowledge of some related ones All our manipulations will be applications of the three basic rules Basic rules apply to any number of varables: P(y) = = = x z P(x, y, z) x P(y x, z)p(x, z) x z P(x y, z)p(y z)p(x, z) P(x z) z (sum rule) (product rule) (Bayes rule)

8 Playing with probabilities Example P(y x, z) = = = = = = P(x, z y)p(y) P(x, z) (Bayes rule) P(x, z y)p(y) P(x z)p(z) (product rule) P(x z, y)p(z y)p(y) P(x z)p(z) (product rule) P(x z, y)p(z, y) P(x z)p(z) (product rule) P(x z, y)p(y z)p(z) P(x z)p(z) (product rule) P(x z, y)p(y z) P(x z)

9 Graphical models Why All probabilistic inference and learning amount at repeated applications of the sum and product rules Probabilistic graphical models are graphical representations of the qualitative aspects of probability distributions allowing to: visualize the structure of a probabilistic model in a simple and intuitive way discover properties of the model, such as conditional independencies, by inspecting the graph express complex computations for inference and learning in terms of graphical manipulations represent multiple probability distributions with the same graph, abstracting from their quantitative aspects (e.g. discrete vs continuous distributions)

10 Bayesian Networks (BN) Description Directed graphical models representing joint distributions p(a, b, c) = p(c a, b)p(a, b) = p(c a, b)p(b a)p(a) a b Fully connected graphs contain no independency assumption Joint probabilities can be decomposed in many different ways by the product rule Assuming an ordering of the variables, a probability distributions over K variables can be decomposed into: c p(x 1,..., x K ) = p(x K x 1,..., x K 1 )... p(x 2 x 1 )p(x 1 )

11 Bayesian Networks Modeling independencies A graph not fully connected contains independency assumptions. The joint probability of a Bayesian network with K nodes is computed as the product of the probabilities of the nodes given their parents: x 1 x 2 x 3 x 4 x 5 x 6 x 7 p(x) = K p(x i pa i ) i=1 The joint probability for the BN in the figure is written as p(x 1 )p(x 2 )p(x 3 )p(x 4 x 1, x 2, x 3 )p(x 5 x 1, x 3 )p(x 6 x 4 )p(x 7 x 4, x 5 )

12 Conditional independence Introduction Two variables a, b are conditionally independent (written a b ) if: p(a, b) = p(a)p(b) Two variables a, b are conditionally independent given c (written a b c ) if: p(a, b c) = p(a c)p(b c) Independency assumptions can be verified by repeated applications of sum and product rules Graphical models allow to directly verify them through the d-separation criterion

13 d-separation Tail-to-tail Joint distribution: p(a, b, c) = p(a c)p(b c)p(c) a and b are not conditionally independent (written a b ): a c b p(a, b) = c p(a c)p(b c)p(c) p(a)p(b) a and b are conditionally independent given c: p(a, b c) = p(a, b, c) p(c) = p(a c)p(b c) c is tail-to-tail wrt to the path a b as it is connected to the tails of the two arrows a c b

14 d-separation Head-to-tail Joint distribution: a c b p(a, b, c) = p(b c)p(c a)p(a) = p(b c)p(a c)p(c) a and b are not conditionally independent: p(a, b) = p(a) c p(b c)p(c a) p(a)p(b) a and b are conditionally independent given c: p(a, b c) = p(b c)p(a c)p(c) p(c) = p(b c)p(a c) c is head-to-tail wrt to the path a b as it is connected to the head of an arrow and to the tail of the other one a c b

15 d-separation Head-to-head Joint distribution: p(a, b, c) = p(c a, b)p(a)p(b) a b a and b are conditionally independent: c p(a, b) = c p(c a, b)p(a)p(b) = p(a)p(b) a and b are not conditionally independent given c: p(a, b c) = p(c a, b)p(a)p(b) p(c) p(a c)p(b c) c is head-to-head wrt to the path a b as it is connected to the heads of the two arrows a c b

16 d-separation General Head-to-head Let a descendant of a node x be any node which can be reached from x with a path following the direction of the arrows A head-to-head node c unblocks the dependency path between its parents if either itself or any of its descendants receives evidence

17 Example of head-to-head connection A toy regulatory network: CPTs Genes A and B have independent prior probabilities: gene value P(value) A active 0.3 A inactive 0.7 gene value P(value) B active 0.3 B inactive 0.7 Gene C can be enhanced by both A and B: A active inactive B B active inactive active inactive C active C inactive

18 Example of head-to-head connection Probability of A active (1) Prior: P(A = 1) = 1 P(A = 0) = 0.3 Posterior after observing active C: P(C = 1 A = 1)P(A = 1) P(A = 1 C = 1) = P(C = 1) Note The probability that A is active increases from observing that its regulated gene C is active

19 Example of head-to-head connection Derivation P(C = 1 A = 1) = P(C = 1, B A = 1) B {0,1} = P(C = 1 B, A = 1)P(B A = 1) B {0,1} = P(C = 1 B, A = 1)P(B) B {0,1} P(C = 1) = P(C = 1, B, A) B {0,1} A {0,1} = B {0,1} A {0,1} P(C = 1 B, A)P(B)P(A)

20 Example of head-to-head connection Probability of A active Posterior after observing that B is also active: P(A = 1 C = 1, B = 1) = P(C = 1 A = 1, B = 1)P(A = 1 B = 1) P(C = 1 B = 1) Note The probability that A is active decreases after observing that B is also active The B condition explains away the observation that C is active The probability is still greater than the prior one (0.3), because the C active observation still gives some evidence in favour of an active A

21 General d-separation criterion d-separation definition Given a generic Bayesian network Given A, B, C arbitrary nonintersecting sets of nodes The sets A and B are d-separated by C if: All paths from any node in A to any node in B are blocked A path is blocked if it includes at least one node s.t. either: the arrows on the path meet tail-to-tail or head-to-tail at the node and it is in C, or the arrows on the path meet head-to-head at the node and neither it nor any of its descendants is in C d-separation implies conditional independency The sets A and B are independent given C ( A B C ) if they are d-separated by C.

22 Example of general d-separation a b c Nodes a and b are not d-separated by c: Node f is tail-to-tail and not observed Node e is head-to-head and its child c is observed a e c f b a b f Nodes a and b are d-separated by f : a f Node f is tail-to-tail and observed e b c

23 Example of i.i.d. samples Maximum-likelihood We are given a set of instances D = {x 1,..., x N } drawn from an univariate Gaussian with unknown mean µ x 1 µ x N All paths between x i and x j are blocked if we condition on µ µ The examples are independent of each other given µ: N p(d µ) = N p(x i µ) i=1 A set of nodes with the same variable type and connections can be compactly represented using the plate notation x n N

24 Markov blanket (or boundary) Definition Given a directed graph with D nodes The markov blanket of node x i is the minimal set of nodes making it x i independent on the rest of the graph: p(x i x j i ) = p(x 1,..., x D ) p(x j i ) = p(x 1,..., x D ) p(x1,..., x D )dx i D k=1 = p(x k pa k ) D k=1 p(x k pa k )dx i All components which do not include x i will cancel between numerator and denominator The only remaining components are: p(x i pa i ) the probability of x i given its parents p(x j pa j ) where pa j includes x i the children of x i with their co-parents

25 Markov blanket (or boundary) d-separation Each parent x j of x i will be head-to-tail or tail-to-tail in the path btw x i and any of x j other neighbours blocked Each child x j of x i will be head-to-tail in the path btw x i and any of x j children blocked Each co-parent x k of a child x j of x i be head-to-tail or tail-to-tail in the path btw x j and any of x k other neighbours blocked x i

26 Markov Networks (MN) or Markov random fields Undirected graphical models Graphically model the joint probability of a set of variables encoding independency assumptions (as BN) Do not assign a directionality to pairwise interactions Do not assume that local interactions are proper conditional probabilities (more freedom in parametrizing them) A C D B

27 Markov Networks (MN) A C B d-separation definition Given A, B, C arbitrary nonintersecting sets of nodes The sets A and B are d-separated by C if: All paths from any node in A to any node in B are blocked A path is blocked if it includes at least one node from C (much simpler than in BN) The sets A and B are independent given C ( A B C ) if they are d-separated by C.

28 Markov Networks (MN) Markov blanket The Markov blanket of a variable is simply the set of its immediate neighbour

29 Markov Networks (MN) Factorization properties We need to decompose the joint probability in factors (like P(x i pa i ) in BN) in a way consistent with the independency assumptions. Two nodes x i and x j not connected by a link are independent given the rest of the network (any path between them contains at least another node) they should stay in separate factors Factors are associated with cliques in the network (i.e. fully connected subnetworks) A possible option is to associate factors only with maximal cliques x 1 x 2 x 3 x 4

30 Markov Networks (MN) Joint probability (1) p(x) = 1 Z Ψ C (x C ) C Ψ C (x C ) : Val(x c ) IR + is a positive function of the value of the variables in clique x C (called clique potential) The joint probability is the (normalized) product of clique potentials for all (maximal) cliques in the network The partition function Z is a normalization constant assuring that the result is a probability (i.e. x p(x) = 1): Z = Ψ C (x C ) x C

31 Markov Networks (MN) Joint probability (2) In order to guarantee that potential functions are positive, they are typically represented as exponentials: Ψ C (x C ) = exp ( E(x c )) E(x c ) is called energy function: low energy means highly probable configuration The joint probability becames a sum of exponentials: ( p(x) = 1 Z exp ) E(x c ) C

32 Markov Networks (MN) Comparison to BN Advantage: More freedom in defining the clique potentials as they don t need to represent a probability distribution Problem: The partition function Z has to be computed summing over all possible values of the variables Solutions: Intractable in general Efficient algorithms exists for certain types of models (e.g. linear chains) Otherwise Z is approximated

33 MN Example: image de-noising Description We are given an image as a grid of binary pixels y i { 1, 1} The image is a noisy version of the original one, where each pixel was flipped with a certain (low) probability The aim is designing a model capable of recovering the clean version

34 MN Example: image de-noising y i x i Markov network We represent the true image as grid of points x i, and connect each point to its observed noisy version y i The MN has two maximal clique types: {x i, y i } for the true/observed pixel pair {x i, x j } for neighbouring pixels The MN has one unobserved non-maximal clique type: the true pixel {x i }

35 MN Example: image de-noising Energy functions We assume low noise the true pixel should usually have same sign as the observed one: E(x i, y i ) = αx i y i α > 0 same sign pixels have lower energy higher probability We assume neighbouring pixels should usually have same sign (locality principle, e.g. background colour): E(x i, x j ) = βx i x j β > 0 We assume the image can have preference for a certain pixel value (either positive or negative): E(x i ) = γx i

36 MN Example: image de-noising Joint probability The joint probability becomes: p(x, y) = 1 Z exp ( E((x, y) C )) (x,y) C = 1 Z exp α x i x j + β x i y i γ i,j i i x i α, β, γ are parameters of the model.

37 MN Example: image de-noising Note all cliques of the same type share the same parameter (e.g. α for cliques {x i, x j }) This parameter sharing (or tying) allows to: reduce the number of parameters, simplifying learning build a network template which will work on images of different sizes.

38 MN Example: image de-noising Inference Assuming parameters are given, the network can be used to produce a clean version of the image: 1 we replace y i with its observed value for each i 2 we compute the most most probable configuration for the true pixels given the observed ones The second computation is complex and can be done in approximate or exact ways (we ll see) with results of different quality.

39 MN Example: image de-noising

40 Relationship between directed and undirected graphs Independency assumptions Directed an undirected graphs are different ways of representing conditional independency assumptions Some assumptions can be perfectly represented in both formalisms, some in only one, some in neither of the two However there is a tight connection between the two It is rather simple to convert a specific directed graphical model into an undirected one (but different directed graphs can be mapped to the same undirected graph) The converse is more complex because of normalization issues

41 From a directed to an undirected model (a) x 1 x 2 x N 1 x N (b) x 1 x 2 x N 1 x N Linear chain The joint probability in the directed model (a) is: p(x) = p(x 1 )p(x 2 x 1 )p(x 3 x 2 ) p(x N x N 1 ) The joint probability in the undirected model (b) is: p(x) = 1 Z ψ 1,2(x 1, x 2 )ψ 2,3 (x 2, x 3 ) ψ N 1,N (x N 1, x N ) In order for the latter to represent the same distribution as the former, we set: ψ 1,2 (x 1, x 2 ) = p(x 1 )p(x 2 x 1 ) ψ 2,3 (x 2, x 3 ) = p(x 3 x 2 ) ψ N 1,N (x N 1, x N ) = p(x N x N 1 ) Z = 1

42 From a directed to an undirected model x 1 x 3 x 1 x 3 x 2 x 2 x 4 x 4 Multiple parents If a node has multiple parents, we need to construct a clique containing all of them, in order for its potential to represent the conditional probability: We add an undirected link between each pair of parents of the node The process is called moralization (marrying the parents! old fashioned..)

43 From a directed to an undirected model Generic graph 1 Build the moral graph connecting the parents in the directed graphs and dropping all arrows 2 initialize all clique potentials to 1 3 for each conditional probability, choose a clique containing all its variables, and multiply its potential with the conditional probability 4 set Z = 1

44 Inference in graphical models Description Assume we have evidence e on the state of a subset of variables in the model E Inference amounts at computing the posterior probability of a subset X of the non-observed variables given the observations: p(x E = e) Note When we need to distinguish between variables and their values, we will indicate random variables with uppercase letters, and their values with lowercase ones.

45 Inference in graphical models Efficiency We can always compute the posterior probability as the ratio of two joint probabilities: p(x E = e) = p(x, E = e) p(e = e) The problem consists of estimating such joint probabilities when dealing with a large number of variables Directly working on the full joint probabilities requires time exponential in the number of variables For instance, if all N variables are discrete and take one of K possible values, a joint probability table has K N entries We would like to exploit the structure in graphical models to do inference more efficiently.

46 Inference in graphical models Inference on a chain (1) p(x) = 1 Z ψ 1,2(x 1, x 2 ) ψ xn 2,x N 1 (x N 2, x N 1 )ψ N 1,N (x N 1, x N ) The marginal probability of an arbitrary x n is: p(x n ) = p(x) x 1 x 2 x n 1 x n+1 x N Only the ψ N 1,N (x N 1, x N ) is involved in the last summation which can be computed first: µ β (x N 1 ) = x N ψ N 1,N (x N 1, x N ) giving a function of x N 1 which can be used to compute the successive summation: µ β (x N 2 ) = x N 1 ψ N 2,N 1 (x N 2, x N 1 )µ β (x N 1 )

47 Inference in graphical models Inference on a chain (2) The same procedure can be applied starting from the other end of the chain, giving µ α (x 2 ) = x 1 ψ 1,2 (x 1, x 2 ) up to µ α (x n ) The marginal probability is thus computed from the normalized contribution coming from both ends as: p(x n ) = 1 Z µ α(x n )µ β (x n ) The partition function Z can now be computed just summing µ α (x n )µ β (x n ) over all possible values of x n

48 Inference in graphical models µ α (x n 1 ) µ α (x n ) µ β (x n ) µ β (x n+1 ) x 1 x n 1 x n x n+1 x N Inference as message passing We can think of µ α (x n ) as a message passing from x n 1 to x n µ α (x n ) = x n 1 ψ xn 1,x n (x n 1, x n )µ α (x n 1 ) We can think of µ β (x n ) as a message passing from x n+1 to x n µ β (x n ) = x n+1 ψ xn,x n+1 (x n, x n+1 )µ β (x n+1 ) Each outgoing message is obtained multiplying the incoming message by the clique potential, and summing over the node values

49 Inference in graphical models Full message passing Suppose we want to know marginal probabilities for a number of different variables x i : 1 We send a message from µ α (x 1 ) up to µ α (x N ) 2 We send a message from µ β (x N ) down to µ β (x 1 ) If all nodes store messages, we can compute any marginal probability as p(x i ) = µ α (x i )µ β (x i ) for any i having sent just a double number of messages wrt a single marginal computation

50 Inference in graphical models Adding evidence If some nodes x e are observed, we simply use their observed values instead of summing over all possible values when computing their messages After normalization by the partition function Z, this gives the conditional probability given the evidence Joint distribution for clique variables It is easy to compute the joint probability for variables in the same clique: p(x n 1, x n ) = 1 Z µ α(x n 1 )ψ xn 1,x n (x n 1, x n )µ β (x n )

51 Inference Directed models The procedure applies straightforwardly to a directed model, replacing clique potentials with conditional probabilities The marginal p(x n ) does not need to be normalized as its already a correct distribution When adding evidence, the message passing procedure computes the joint probability of the variable and the evidence, and it has to be normalized to obtain the conditional probability: p(x n X e = x e ) = p(x n, X e = x e ) p(x e = x e ) = p(x n, X e = x e ) X n p(x n, X e = x e )

52 Inference Inference on trees (a) (b) (c) Efficient inference can be computed for the broaded family of tree-structured models: undirected trees (a) undirected graphs with a single path for each pair of nodes directed trees (b) directed graphs with a single node (the root) with no parents, and all other nodes with a single parent directed polytrees (c) directed graphs with multiple parents for node and multiple roots, but still as single (undirected) path between each pair of nodes

53 Factor graphs Description Efficient inference algorithms can be better explained using an alternative graphical representation called factor graph A factor graph is a graphical representation of a (directed or undirected) graphical model highlighting its factorization (conditional probabilities or clique potentials) The factor graph has one node for each node in the original graph The factor graph has one additional node (of a different type) for each factor A factor node has undirected links to each of the node variables in the factor

54 Factor graphs: examples x 1 x 2 x 1 x 2 f x 1 x 2 f c f a f b x 3 x 3 x 3 p(x3 x1,x2)p(x1)p(x2) f(x1,x2,x3)=p(x3 x1,x2)p(x1)p(x2) fc(x1,x2,x3)=p(x3 x1,x2) fa(x1)=p(x1) fb(x2)=p(x2)

55 Factor graphs: examples x 1 x 2 x 1 x 2 f x 1 x 2 f a f b x 3 x 3 x 3 Note f is associated to the whole clique potential f a is associated to a maximal clique f b is associated to a non-maximal clique

56 Inference The sum-product algorithm The sum-product algorithm is an efficient algorithm for exact inference on tree-structured graphs It is a message passing algorithm as its simpler version for chains We will present it on factor graphs, thus unifying directed and undirected models We will assume a tree-structured graph, giving rise to a factor graph which is a tree

57 Inference Computing marginals We want to compute the marginal probability of x: p(x) = p(x) x\x Generalizing the message passing scheme seen for chains, this can be computed as the product of messages coming from all neighbouring factors f s : p(x) = 1 Z s ne(x) µ fs x(x) Fs(x, Xs) f s µ fs x(x) x

58 Inference x M µ xm f s (x M ) f s µ fs x(x) x Factor messages x m G m (x m, X sm ) Each factor message is the product of messages coming from nodes other than x, times the factor, summed over all possible values of the factor variables other than x (x 1,..., x M ): µ fs x(x) = x 1 x M f s (x, x 1,..., x M ) m ne(f s)\x µ xm f s (x m )

59 Inference f L x m f s f l F l (x m, X ml ) Node messages Each message from node x m to factor f s is the product of the factor messages to x m coming from factors other than f s : µ xm fs (x m ) = µ fl x m (x m ) l ne(x m)\f s

60 Inference Initialization Message passing starts from leaves, either factors or nodes Messages from leaf factors are initialized to the factor itself (there will be no x m different from the destination on which to sum over) µ f x (x) = f(x) f x Messages from leaf nodes are initialized to 1 µ x f (x) = 1 x f

61 Inference Message passing scheme The node x whose marginal has to be computed is designed as root. Messages are sent from all leaves to their neighbours Each internal node sends its message towards the root as soon as it received messages from all other neighbours Once the root has collected all messages, the marginal can be computed as the product of them

62 Inference Full message passing scheme In order to be able to compute marginals for any node, messages need to pass in all directions: 1 Choose an arbitrary node as root 2 Collect messages for the root starting from leaves 3 Send messages from the root down to the leaves All messages passed in all directions using only twice the number of computations used for a single marginal

63 Inference example x 1 x 2 x 3 f a f b f c x 4 Consider the unnormalized distribution p(x) = f a (x 1, x 2 )f b (x 2, x 3 )f c (x 2, x 4 )

64 Inference example x 1 x 2 x 3 f a f b f c x 4 Choose x 3 as root

65 Inference example x 1 x 2 x 3 x 4 Send initial messages from leaves µ x1 f a (x 1 ) = 1 µ x4 f c (x 4 ) = 1

66 Inference example x 1 x 2 x 3 x 4 Send messages from factor nodes to x 2 µ fa x 2 (x 2 ) = x 1 f a (x 1, x 2 ) µ fc x 2 (x 2 ) = x 4 f c (x 2, x 4 )

67 Inference example x 1 x 2 x 3 x 4 Send message from x 2 to factor node f b µ x2 f b (x 2 ) = µ fa x 2 (x 2 )µ fc x 2 (x 2 )

68 Inference example x 1 x 2 x 3 x 4 Send message from f b to x 3 µ fb x 3 (x 3 ) = x 2 f b (x 2, x 3 )µ x2 f b (x 2 )

69 Inference example x 1 x 2 x 3 x 4 Send message from root x 3 µ x3 f b (x 3 ) = 1

70 Inference example x 1 x 2 x 3 x 4 Send message from f b to x 2 µ fb x 2 (x 2 ) = x 3 f b (x 2, x 3 )

71 Inference example x 1 x 2 x 3 x 4 Send messages from x 2 to factor nodes µ x2 f a (x 2 ) = µ fb x 2 (x 2 )µ fc x 2 (x 2 ) µ x2 f c (x 2 ) = µ fb x 2 (x 2 )µ fa x 2 (x 2 )

72 Inference example x 1 x 2 x 3 x 4 Send messages from factor nodes to leaves µ fa x 1 (x 1 ) = x 2 f a (x 1, x 2 )µ x2 f a (x 2 ) µ fc x 4 (x 4 ) = x 2 f c (x 2, x 4 )µ x2 f c (x 2 )

73 Inference example x 1 x 2 x 3 f a f b f c x 4 Compute for instance the marginal for x 2 p(x 2 ) = µ fa x2 (x 2 )µ fb x 2 (x 2 )µ fc x2 (x 2 ) [ ] [ ] [ ] = f a (x 1, x 2 ) f b (x 2, x 3 ) f c (x 2, x 4 ) x 1 x 3 x 4 = f a (x 1, x 2 )f b (x 2, x 3 )f c (x 2, x 4 ) x 1 x 3 x 4 = p(x) x 1 x 3 x 4

74 Inference Normalization p(x) = 1 Z s ne(x) µ fs x(x) If the factor graph came from a directed model, a correct probability is already obtained as the product of messages (Z = 1) If the factor graph came from an undirected model, we simply compute the normalization Z summing the product of messages over all possible values of the target variable: Z = x s ne(x) µ fs x(x)

75 Inference Factor marginals p(x s ) = 1 Z f s(x s ) m ne(f s) µ xm f s (x m ) factor marginals can be computed from the product of neighbouring nodes messages and the factor itself Adding evidence If some nodes x e are observed, we simply use their observed values instead of summing over all possible values when computing their messages After normalization by the partition function Z, this gives the conditional probability given the evidence

76 Inference example A C D A P(A) C P(C D) D P(D) B P(B A,C)P(A)P(C D)P(D) P(B A,C) B Directed model Take a directed graphical models Build a factor graph representing it Compute the marginal for a variable (e.g. B)

77 Inference example A C D A P(A) C P(C D) D P(D) B P(B A,C)P(A)P(C D)P(D) P(B A,C) B Compute the marginal for B Leaf factor nodes send messages: µ fa A(A) = P(A) µ fd D(D) = P(D)

78 Inference example A C D A P(A) C P(C D) D P(D) B P(B A,C)P(A)P(C D)P(D) P(B A,C) B Compute the marginal for B A and D send messages: µ A fa,b,c (A) = µ fa A(A) = P(A) µ D fc,d (D) = µ fd D(D) = P(D)

79 Inference example A C D A P(A) C P(C D) D P(D) B P(B A,C)P(A)P(C D)P(D) P(B A,C) B Compute the marginal for B f C,D sends message: µ fc,d C(C) = D P(C D)µ fd D(D) = D P(C D)P(D)

80 Inference example A C D A P(A) C P(C D) D P(D) B P(B A,C)P(A)P(C D)P(D) P(B A,C) B Compute the marginal for B C sends message: µ C fa,b,c (C) = µ fc,d C(C) = D P(C D)P(D)

81 Inference example A C D A P(A) C P(C D) D P(D) B P(B A,C)P(A)P(C D)P(D) P(B A,C) B Compute the marginal for B f A,B,C sends message: µ fa,b,c B(B) = P(B A, C)µ C fa,b,c (C)µ A fa,b,c (A) A C = P(B A, C)P(A) P(C D)P(D) A D C

82 Inference example A C D A P(A) C P(C D) D P(D) B P(B A,C)P(A)P(C D)P(D) P(B A,C) B Compute the marginal for B The desired marginal is obtained: P(B) = µ fa,b,c B(B) = P(B A, C)P(A) A C D = P(B A, C)P(A)P(C D)P(D) A C D = P(A, B, C, D) A D C P(C D)P(D)

83 Inference example A C D A P(A) C P(C D) D P(D) B P(B A,C)P(A)P(C D)P(D) P(B A,C) B Compute the joint for A, B, C Send messages to factor node f A,B,C P(A, B, C) = f A,B,C (A, B, C) µ A fa,b,c (A) µ C fa,b,c (C) µ B fa,b,c (B) = P(B A, C) P(A) P(C D)P(D) 1 = P(B A, C)P(A)P(C D)P(D) D D = D P(A, B, C, D)

84 Inference Finding the most probable configuration Note Given a joint probability distribution p(x) We wish to find the configuration for variables x having the highest probability: for which the probability is: x max = argmax x p(x) p(x max ) = max p(x) x We want the configuration which is jointly maximal for all variables We cannot simply compute p(x i ) for each i (using the sum-product algorithm) and maximize it

85 Inference The max-product algorithm p(x max ) = max x p(x) = max x 1 max p(x) x M As for the sum-product algorithm, we can exploit the distribution factorization to efficiently compute the maximum It suffices to replace sum with max in the sum-product algorithm Linear chain max x p(x) = max x 1 = 1 Z max x 1 [ ] 1 max x N Z ψ 1,2(x 1, x 2 ) ψ N 1,N (x N 1, x N ) [ [ ]] ψ 1,2 (x 1, x 2 ) max ψ N 1,N (x N 1, x N ) x N

86 Inference Message passing As for the sum-product algorithm, the max-product can be seen as message passing over the graph. The algorithm is thus easily applied to tree-structured graphs via their factor trees: µ f x (x) = max f (x, x 1,..., x M ) x 1,...x M µ x f (x) = µ fl x(x) l ne(x)\f m ne(f )\x µ xm f (x m )

87 Inference Recoving maximal configuration Messages are passed from leaves to an arbitrarily chosen root x r The probability of maximal configuration is readily obtained as: p(x max ) = max x r l ne(x r ) µ fl x r (x r ) The maximal configuration for the root is obtained as: = argmax xr µ fl x r (x r ) x max r l ne(x r ) We need to recover maximal configuration for the other variables

88 Inference Recoving maximal configuration When sending a message towards x, each factor node should store the configuration of the other variables which gave the maximum: φ f x (x) = argmax x1,...,x M f (x, x 1,..., x M ) m ne(f )\x µ xm f (x m ) When the maximal configuration for the root node x r has been obtained, it can be used to retrieve the maximal configuration for the variables in neighbouring factors from: x1 max,..., xm max = φ f x r (xr max ) The procedure can be repeated back-tracking to the leaves, retrieving maximal values for all variables

89 Recoving maximal configuration Example for linear chain xn max = argmax xn µ fn 1,N x N (x N ) xn 1 max = φ f N 1,N x N (xn max ) xn 2 max = φ f N 2,N 1 x N 1 (xn 1 max ). x max 1 = φ f1,2 x 2 (x max 2 )

90 Recoving maximal configuration k = 1 k = 2 k = 3 n 2 n 1 n n + 1 Trellis for linear chain A trellis or lattice diagram shows the K possible states of each variable x n one per row For each state k of a variable x n, φ fn 1,n x n (x n ) defines a unique (maximal) previous state, linked by an edge in the diagram Once the maximal state for the last variable x N is chosen, the maximal states for other variables are recovering following the edges backward.

91 Inference Underflow issues The max-product algorithm relies on products (no summation) Products of many small probabilities can lead to underflow problems This can be addressed computing the logarithm of the probability instead The logarithm is monotonic, thus the proper maximal configuration is recovered: ( ) log max p(x) = max log p(x) x x The effect is replacing products with sums (of logs) in the max-product algorithm, giving the max-sum one

92 Inference Instances of inference algorithms The sum-product and max-product algorithms are generic algorithms for exact inference in tree-structured models. Inference algorithms for certain specific types of models can be seen as special cases of them We will see examples (forward-backward as sum-product and Viterbi as max-product) for Hidden Markov Models and Conditional Random Fields

93 Inference Exact inference on general graphs The sum-product and max-product algorithms can be applied to tree-structured graphs Many pratical application require graphs with loops An extension of this algorithms to generic graphs can be achieved with the junction tree algorithm The algorithm does not work on factor graphs, but on junction trees, tree-structured graphs with nodes containing clusters of variables of the original graph A message passing scheme analogous to the sum-product and max-product algorithms is run on the junction tree Problem The complexity on the algorithm is exponential on the maximal number of variables in a cluster, making it intractable for large complex graphs.

94 Inference Approximate inference In cases in which exact inference is intractable, we resort to approximate inference techniques A number of techniques for approximate inference exist: loopy belief propagation message passing on the original graph even if it contains loops variational methods deterministic approximations, assuming the posterior probability (given the evidence) factorizes in a particular way sampling methods approximate posterior is obtained sampling from the network

95 Inference Loopy belief propagation Apply sum-product algorithm even if it is not guaranteed to provide an exact solution We assume all nodes are in condition of sending messages (i.e. they already received a constant 1 message from all neighbours) A message passing schedule is chosen in order to decide which nodes start sending messages (e.g. flooding, all nodes send messages in all directions at each time step) Information flows many times around the graph (because of the loops), each message on a link replaces the previous one and is only based on the most recent messages received from the other neighbours The algorithm can eventually converge (no more changes in messages passing through any link) depending on the specific model over which it is applied

96 Approximate inference Sampling methods Given the joint probability distribution p(x) A sample from the distribution is an instantiation of all the variables X according to the probability p. Samples can be used to approximate the probability of a certain assignment Y = y for a subset Y X of the variables: We simply count the fraction of samples which are consistent with the desired assignment We usually need to sample from a posterior distribution given some evidence E = e

97 Sampling methods Markov Chain Monte Carlo (MCMC) We usually cannot directly sample from the posterior distribution (too expensive) We can instead build a random process which gradually samples from distributions closer and closer to the posterior The state of the process is an instantiation of all the variables The process can be seen as randomly traversing the graph of states moving from one state to another with a certain probability. After enough time, the probability of being in any particular state is the desired posterior. The random process is a Markov Chain

98 Learning graphical models Parameter estimation We assume the structure of the model is given We are given a dataset of examples D = {x(1),..., x(n)} Each example x(i) is a configuration for all (complete data) or some (incomplete data) variables in the model We need to estimate the parameters of the model (CPD or weights of clique potentials) from the data The simplest approach consists of learning the parameters maximizing the likelihood of the data: θ max = argmax θ p(d θ) = argmax θ L(D, θ)

99 Learning Bayesian Networks Θ1 X1(1) X1(2) X1(N) X2(1) X3(1) X2(2) X3(2) X2(N) X3(N) Θ2 1 Θ3 1 Maximum likelihood estimation, complete data p(d θ) = = N p(x(i) θ) i=1 N i=1 j=1 m p(x j (i) pa j (i), θ) examples independent given θ factorization for BN

100 Learning Bayesian Networks Θ1 X1(1) X1(2) X1(N) X2(1) X3(1) X2(2) X3(2) X2(N) X3(N) Θ2 1 Θ3 1 Maximum likelihood estimation, complete data N m p(d θ) = p(x j (i) pa j (i), θ) i=1 j=1 N m = p(x j (i) pa j (i), θ Xj pa j ) i=1 j=1 factorization for BN disjoint CPD parameters

101 Learning graphical models Maximum likelihood estimation, complete data The parameters of each CPD can be estimated independently: θ max X j Pa j = argmax θxj Pa j N p(x j (i) pa j (i), θ Xj Pa j ) i=1 } {{ } L(θ Xj Pa,D) j A discrete CPD P(X U), can be represented as a table, with: a number of rows equal to the number Val(X) of configurations for X a number of columns equal to the number Val(U) of configurations for its parents U each table entry θ x u indicating the probability of a specific configuration of X = x and its parents U = u

102 Learning graphical models Maximum likelihood estimation, complete data Replacing p(x(i) pa(i)) with θ x(i) u(i), the local likelihood of a single CPD becames: L(θ X Pa, D) = = N p(x(i) pa(i), θ X Paj ) i=1 N i=1 = θ x(i) u(i) u Val(U) x Val(X) θ Nu,x x u where N u,x is the number of times the specific configuration X = x, U = u was found in the data

103 Learning graphical models Maximum likelihood estimation, complete data A column in the CPD table contains a multinomial distribution over values of X for a certain configuration of the parents U Thus each column should sum to one: x θ x u = 1 Parameters of different columns can be estimated independently For each multinomial distribution, zeroing the gradient of the maximum likelihood and considering the normalization constraint, we obtain: θx u max = N u,x x N u,x The maximum likelihood parameters are simply the fraction of times in which the specific configuration was observed in the data

104 Learning Bayesian Networks Example Training examples as (A, B, C) tuples: (act,act,act),(act,inact,act), (act,inact,act),(act,inact,inact), (inact,act,act),(inact,act,inact), (inact,inact,inact),(inact,inact,inact), (inact,inact,inact),(inact,inact,inact), (inact,inact,inact),(inact,inact,inact). Fill CPTs with counts

105 Learning Bayesian Networks Example Training examples as (A, B, C) tuples: (act,act,act),(act,inact,act), (act,inact,act),(act,inact,inact), (inact,act,act),(inact,act,inact), (inact,inact,inact),(inact,inact,inact), (inact,inact,inact),(inact,inact,inact), (inact,inact,inact),(inact,inact,inact). Fill CPTs with counts gene value counts A active 4 A inactive 8 gene value counts B active 3 B inactive 9

106 Learning Bayesian Networks Example Training examples as (A, B, C) tuples: (act,act,act),(act,inact,act), (act,inact,act),(act,inact,inact), (inact,act,act),(inact,act,inact), (inact,inact,inact),(inact,inact,inact), (inact,inact,inact),(inact,inact,inact), (inact,inact,inact),(inact,inact,inact). Normalize counts columnwise gene value counts A active 4/12 A inactive 8/12 gene value counts B active 3/12 B inactive 9/12

107 Learning Bayesian Networks Example Training examples as (A, B, C) tuples: (act,act,act),(act,inact,act), (act,inact,act),(act,inact,inact), (inact,act,act),(inact,act,inact), (inact,inact,inact),(inact,inact,inact), (inact,inact,inact),(inact,inact,inact), (inact,inact,inact),(inact,inact,inact). Normalize counts columnwise gene value counts A active 0.33 A inactive 0.67 gene value counts B active 0.25 B inactive 0.75

108 Learning Bayesian Networks Example Training examples as (A, B, C) tuples: (act,act,act),(act,inact,act), (act,inact,act),(act,inact,inact), (inact,act,act),(inact,act,inact), (inact,inact,inact),(inact,inact,inact), (inact,inact,inact),(inact,inact,inact), (inact,inact,inact),(inact,inact,inact). Fill CPTs with counts A active inactive B B active inactive active inactive C active C inactive

109 Learning Bayesian Networks Example Training examples as (A, B, C) tuples: (act,act,act),(act,inact,act), (act,inact,act),(act,inact,inact), (inact,act,act),(inact,act,inact), (inact,inact,inact),(inact,inact,inact), (inact,inact,inact),(inact,inact,inact), (inact,inact,inact),(inact,inact,inact). Fill CPTs with counts A active inactive B B active inactive active inactive C active C inactive

110 Learning Bayesian Networks Example Training examples as (A, B, C) tuples: (act,act,act),(act,inact,act), (act,inact,act),(act,inact,inact), (inact,act,act),(inact,act,inact), (inact,inact,inact),(inact,inact,inact), (inact,inact,inact),(inact,inact,inact), (inact,inact,inact),(inact,inact,inact). Normalize counts columnwise A active inactive B B active inactive active inactive C active 1/1 2/3 1/2 0/6 C inactive 0/1 1/3 1/2 6/6

111 Learning Bayesian Networks Example Training examples as (A, B, C) tuples: (act,act,act),(act,inact,act), (act,inact,act),(act,inact,inact), (inact,act,act),(inact,act,inact), (inact,inact,inact),(inact,inact,inact), (inact,inact,inact),(inact,inact,inact), (inact,inact,inact),(inact,inact,inact). Normalize counts columnwise A active inactive B B active inactive active inactive C active C inactive

112 Learning graphical models Adding priors ML estimation tends to overfit the training set Configuration not appearing in the training set will receive zero probability A common approach consists of combining ML with a prior probability on the parameters, achieving a maximum-a-posteriori estimate: θ max = argmax θ p(d θ)p(θ)

113 Learning graphical models Dirichlet priors The conjugate (read natural) prior for a multinomial distribution is a Dirichlet distribution with parameters α x u for each possible value of x The resulting maximum-a-posteriori estimate is: x u = N u,x + α x u ( ) Nu,x + α x u θ max x The prior is like having observed α x u imaginary samples with configuration X = x, U = u

114 Learning Bayesian Networks Example Training examples as (A, B, C) tuples: (act,act,act),(act,inact,act), (act,inact,act),(act,inact,inact), (inact,act,act),(inact,act,inact), (inact,inact,inact),(inact,inact,inact), (inact,inact,inact),(inact,inact,inact), (inact,inact,inact),(inact,inact,inact). Fill CPTs with imaginary counts (priors) A active inactive B B active inactive active inactive C active C inactive

115 Learning Bayesian Networks Example Training examples as (A, B, C) tuples: (act,act,act),(act,inact,act), (act,inact,act),(act,inact,inact), (inact,act,act),(inact,act,inact), (inact,inact,inact),(inact,inact,inact), (inact,inact,inact),(inact,inact,inact), (inact,inact,inact),(inact,inact,inact). Fill CPTs with imaginary counts (priors) A active inactive B B active inactive active inactive C active C inactive

116 Learning Bayesian Networks Example Training examples as (A, B, C) tuples: (act,act,act),(act,inact,act), (act,inact,act),(act,inact,inact), (inact,act,act),(inact,act,inact), (inact,inact,inact),(inact,inact,inact), (inact,inact,inact),(inact,inact,inact), (inact,inact,inact),(inact,inact,inact). Add counts from training data A active inactive B B active inactive active inactive C active C inactive

117 Learning Bayesian Networks Example Training examples as (A, B, C) tuples: (act,act,act),(act,inact,act), (act,inact,act),(act,inact,inact), (inact,act,act),(inact,act,inact), (inact,inact,inact),(inact,inact,inact), (inact,inact,inact),(inact,inact,inact), (inact,inact,inact),(inact,inact,inact). Normalize counts columnwise A active inactive B B active inactive active inactive C active 2/3 3/4 2/4 1/8 C inactive 1/3 2/4 2/4 7/8

118 Learning Bayesian Networks Example Training examples as (A, B, C) tuples: (act,act,act),(act,inact,act), (act,inact,act),(act,inact,inact), (inact,act,act),(inact,act,inact), (inact,inact,inact),(inact,inact,inact), (inact,inact,inact),(inact,inact,inact), (inact,inact,inact),(inact,inact,inact). Normalize counts columnwise A active inactive B B active inactive active inactive C active C inactive

119 Learning graphical models Incomplete data With incomplete data, some of the examples miss evidence on some of the variables Counts of occurrences of different configurations cannot be computed if not all data are observed The full Bayesian approach of integrating over missing variables is often intractable in practice We need approximate methods to deal with the problem

120 Learning with missing data: Expectation-Maximization E-M for Bayesian nets in a nutshell Sufficient statistics (counts) cannot be computed (missing data) Fill-in missing data inferring them using current parameters (solve inference problem to get expected counts) Compute parameters maximizing likelihood (or posterior) of such expected counts Iterate the procedure to improve quality of parameters

121 Learning in Markov Networks Maximum likelihood estimation For general MN, the likelihood function cannot be decomposed into independent pieces for each potential because of the global normalization (Z ) L(E, D) = N i=1 ( 1 Z exp C E c (x c (i)) However the likelihood is concave in E C, the energy functionals whose parameters have to be estimated The problem is an unconstrained concave maximization problem solved by gradient ascent (or second order methods) )

122 Learning in Markov Networks Maximum likelihood estimation For each configuration u c Val(x c ) we have a parameter E c,uc IR (ignoring parameter tying) The partial derivative of the log-likelihood wrt E c,uc is: log L(E, D) E c,uc = N [δ(x c (i), u c ) p(x c = u c E)] i=1 = N uc N p(x c = u c E) The derivative is zero when the counts of the data correspond to the expected counts predicted by the model In order to compute p(x c = u c E), inference has to be performed on the Markov network, making learning quite expensive

123 Learning structure of graphical models Approaches constraint-based test conditional independencies on the data and construct a model satisfying them score-based assign a score to each possible structure, define a search procedure looking for the structure maximizing the score model-averaging assign a prior probability to each structure, and average prediction over all possible structures weighted by their probabilities (full Bayesian, intractable)